Automated Machine Learning danger

Thanks to machine learning, computers this month learnt to be racist Nazis and to win at Go.  In under 24 hours Microsoft’s Tay chatbot was born, converted by Twitter users into a racist anti-feminist Nazi and turned off (Hacker News).  As so often in computing the quality of the output is a derivative of the quality of the input…

A week earlier machine learning and pattern recognition cracked the game of go – reported here on Wired.

This required 1,920 CPUs and 280 GPUs according to Economist report
Don’t be fooled by the relatively ‘low’ number of GPUs: each GPU runs thousands of parallel calculation threads on Arithmetic Logic Units – simpler than general purpose CPU and so an order of magnitude more efficient in terms of power consumption/heat dissipation etc:  for example the nVidia “Visual Computing Appliance” combines 8 GPUs to run 24,576 threads accessing 256GB of memory.

These GPUs power the convolutional neural networks capable of pattern and image recognition and the machine learning which differentiate this generation of Artificial Intelligence. Go and real-life are too complex to calculate simply by the raw processing power of 1,920CPUs, a combination of best-guess approximations and learned techniques are needed.

While AlphaGo is reported as a Google product, it is actually developed in London by an international team led by Dennis Hassabis and Deep Mind – bought by Google in 2014 – see Guardian article, good to see London still a centre for innovation.



Leave a Reply

Please log in using one of these methods to post your comment: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s